I want to write a TypeScript type based on a union such that each member of the union is transformed a certain way if it matches a certain shape, and is otherwise discarded. In working on this, I found that in TypeScript, the result of applying a type to a generic is different than substituting that type into the body of the generic in place of the type parameter. For example:

type T = "a" | "b" | "c" | 1 | 2 | 3;
// type U = never
type U = T extends string ? T : never;
type GV<GT> = GT extends string ? GT : never;
// type V = "a" | "b" | "c"
type V = GV<T>;

Type U and V are different, even though substituting T into the body of GV in place of GT would yield exactly the definition of U.

Given that referential transparency is not an accurate way to think about TypeScript generics, what is the intended way to think about them? Where is this behavior documented? I couldn’t find any information about it in the official documentation for generics.